AI in videogames is about to get really good

Am I the only one that's really excited for the onset of Artificial Intelligence? Even GPU companies are starting to build GPU's that are designed to do tasks that help complex AI function.

Think of how interesting videogames are going to get once they the AI in them is able to think in a much more complex manner.

Other urls found in this thread:

deepmind.com/blog/deepmind-and-blizzard-release-starcraft-ii-ai-research-environment/
sscaitournament.com/
twitter.com/NSFWRedditGif

I've said this before but good AI is bad for games.
To have tight gameplay you need a predictable AI

I think it entirely depends on what the AI is being used for. If the AI is for people it should be able to do anything to simulate a human within the limits of the game. That would make for some interesting and challenging gameplay.

I remember hearing the same fucking thing when 7th Gen was about to be released and yet we got even shittier A.I due to online focus and pure kikery.

I'm not talking about what most of us are used to, I'm talking about legit supercomputer AI being used in games of the future. Programs that can change what they do (within the limits of the game)

I am going to assume what you mean is that it's better to have a game that is predictable so the player can make meaningful decisions easier, and not somebody that thinks AI should be a pushover.

Are Fear and Halo CE/Reach AI the only FPS games with actuall good AI

I do mean the first, but I still like the idea of having a pushover AI

Go back to CoD laddeh

other guy may be a pleb, but you are literally abnormal

We're obviously talking about A.I used ingame you goddamn nigger, of course a well scripted A.I with with good pathfinding/colision/aim and plenty of variables is what we're talking about, not a potencial Skynet wannabe.

In what board do you think you are?
Get the fuck out of here with your bait nigger

what the fuck would that do
if you want to tell the guy to kill himself just say it

you have to program an AI to play a video game. Complex AI in a game is just more "if" statements to be processed. That's all it is in games with good AI.

It's highly unlikely in our lifetimes we will see an AI that could learn to play a video game by itself without any training.

That's not to mention the AI would have to behave like an ingame character and not as a player.

While it would be cool to train an AI to play Quake indistinguishably from an actual player it's unlikely you'd be able to just download such a thing and play with it within our lifetimes.

The main reason AI sucks in games nowadays is largely because players don't want the game to be too hard and for stealth to be really punishing. It's why enemies in games like MGS5 can't see you unless you're within 5 feet of them.

STALKER has pretty good AI. Half-Life has pretty good AI. Unreal has amazing bot AI. There's plenty of games with solid AI that does things to adapt to the player.

No

AI that learns from you is more interesting than AI that is pre-programmed, but it can lead to bizarre shit sometimes. And the problem is that AI that is actually challenging will only ever come out for games tailored specifically for streambait.

Not happening as long as consoles running at half or less the clock rate of decent PCs exist.

It's already been done. Three years ago.

FEAR AI was actually retarded very simple, though. The level design was just so good that you wouldn't notice on your first play-through.

It was both you nigger
The level design was simple but extremely well though out and the A.I was extremely well scripted.

This is just it learning how to move in 4 directions and press A or B, and learns when to press the button after failing several times.

Once you introduce things like actual three dimensions and other players the AI wandering into a corner for an hour before finally figuring out what direction another player is isn't going to be enjoyable to play against.

Most of you niggers are arguing in different directions and then getting upset you reach different results.

A complex AI can still be predictable if it makes sensible choices, things you can predict with a simple "what would I do in it's position?"
The problem is that today most games don't even follow that because the AI and the levels are very limited and they get 3-5 actions they can opt for and that's it.

Even having some randomness applied to it doesn't make it too bad, You can know overall what the AI would pick with a few surprises to keep you on your toes every now and then.

Personnaly, I'm expecting simulations of villages to advance considerably to the point you can have an RPG set in a world where economy and politics are proper battlefields where the AI can compete with very long and detailed plans for long term.

These niggers can't even be arsed to optimize their games properly and inflate required specs because they know autists will keep upgrading their PCs anyway.

This isn't true but you can definitely have good enemy design without good AI. Doom being a great example since the enemies fill so many different roles.

Completely unrelated to the thread but Prinny is a great character

Even more unrelated but I hope Roll wins the next animation poll

...

...

That's prinny hot, dood!

AI is simply as good as the programmer is willing to put in and adaptive AI will always inevitably cheat.
I remember this shit about college kids programming an AI to play AoEII and all it fucking did was punch in black death frame perfectly inevitably in all tests of it.

I don't want some random japanese porn game AI to gain sentience and exterminate humanity.

you say video games are going to start focusing on singleplayer aspects? im not so sure about that

A good AI, like any feature, needs to be programmed well. With the current cucked industry full of clueless normalfags we won't be seeing good AI for a long time.

Maybe selfemployed autists like Toady can do something with it. Big AAA studios? Never. Civ 6 has WORSE AI than Civ 5 for example.

if you want rolls go check out the stuff by sb

did they ever improve civ 6's ai?

...

All they did was 2 patches with a third on the way only now. Game is basically dead.

People still do this?

Who said I bought the game?
Also DLC won't fix it.

I never said you did
But the game sold like 1.5Mil

Yeah complete bullshit when they can calculate probabilities faster and better than humans you'll die before the game even loads

Nonsense, its because most young developers don't know how to properly program and build everything in prefabricated engines. Its not that the gamer is lazy, its the developers are lazy. This issue is not even one that the publishers can be blamed for. Bad AI is 200% on the shoulder of the shitty Developers.

Yeah I'm hearing this since 2001 and it never did, in fact it got worse.

In my lifetime, the only AI advancements I hope for are in writing, interactivity, and image creation.

VR is decades away from being viable, so let's be realistic here. The best thing we can hope for from AI is for it to be a comprehensive porn creation tool. Something you can train to make comics exactly tailored to your liking and fetishes. This is a realistic, obtainable goal. If conversational, creative AI comes far enough, you might get AI that can ERP with you on demand.

The far-fetched hopes for AI would be something like an MMO where the NPCs are all AI-run and have their own personalities, needs, and so on, so the world feels more lively and active.

The distant future might hold promise of full sensory immersion, brain-jacked vidya where you can tell an AI, holodeck style, exactly what you wish to see and do, and it will create a game for you, but that's a long way off. For now, I'll settle for a computer that can typefuck for me.

AI in vidya will never get good, stupid is much easier and cheaper to get a good end result with
We've had pretty good chatbots for fucking years now but the only game that used the technology was Event[0], and even though that shit just came out last year, they coded their own from scratch and it was scripted to all hell

Well that entirely depends on what market you're talking about. Are you talking about indies mashing together a game in Unity? Then yes, shit AI is probably them being inept. Are you talking about AAA games with custom engines from scratch? Then it's 90% management/marketing restricting game difficulty, often forcing devs to butcher their AI in the process.

This is literally the only thing about AI left that needs to be touched by any competent and respected developer. The AI acts based on its current needs, not only reacting in a scripted way or whatever because x-event happened. Give it muliple goals and weigh them, and you have a system that can work in nearly any game genre. FPS bots and tactical AI would be a hell of a lot better than the shit we have now, open world AI can become much more reactive and believeable than the basic ass shit we have now, fighting game AI wouldn't have to cheat anymore to win, and a bunch of other shit

Isn't that just called "playing video games?"


And then we can put it in waifubots and video games will be made obsolete.

...

Tynan pls, go finish Rimworld you lazy fuck

I dunno, Op. A spammy and overly defensive AI that can react to your inputs is better than one who learns from player input. Specially since some players are panicky and stupid and don't know half of what they are doing and the others will just willingly jump off ledges to ruin the AI and teach it something it should not learn.

Focus groups shat on good AI over a decade ago. Good AI will exist, but only in niche games. Then some AAA like TLoU will have enemies that can flank and get called Citizen Kane again.

Like said. It's not about the AI being good but rather the AI fitting with the gameplay style. Doing predictable things yet sometimes surprising you with its behavior is most important.

Would you buy it?

AI in videogames will not get good because of two simple facts:
It is not easily marketable and it takes effort to create.

I'd take AI that learns from results instead of from what the player does any day of the week.
It might seem like it doesn't count but boss characters in Mortal Kombat doing input reading never feel fair at all, however AI in Guilty Gear Arcade mode actually adapts to how you fight and the tactics that worked for one fight might not work for the next one, forcing you to change your game.


No. Same problem as (one of the many) problems VR has: it's too gimmicky.
Just consider at what price it would have to be sold to cover it's costs and turn up a profit, then consider what you could do with it that justifies the price.

Might as well buy an actual robot\android that uses it instead of limiting it to a PC or videogames.


Oh come the fuck on now. It's the thing people complain the most for 4X and GSG games, that the AI is simple atrocius and never puts up a fair fight. Even for most RTS (not all of them), the harder AIs need a resource\production boost to stay competitive and nobody likes that shit since it's like playing against someone with an handicap.

And this isn't even considering simulators and RPG games like SimCity, Banished, any Tycoon game and basically every single FPS that tried to have an open world with NPCs having their own routine.
Every single one of these games have incredibly simple AI that's not very interesting on an individual level but with improved AI you could have something far more interesting done with those concepts.

>no good h-game AI yet
super gay


Literally Quake 3.

Both genres nobody plays.
Dead genre because nobody played them.

A game I've been playing lately is day of infamy, it includes a single player mode with bots as "practice".

Trouble is the AI is too fucking good, you get sniped from halfway across the map and your team's AI is a bunch of dullards to compensate for your human advantage. What this means is the actual online multiplayer is fucking easier than bot mode. It also doesn't help that DoI has this system where you just shift into a bot rather than respawning, causing nothing but cheap ass deaths.

Rising Storm did bots really well, they're essentially cannon fodder while you're really supposed to worry about other players. The bots offer just the right amount of challenge and assist in your immersion. You need to make the AI stupid enough so that players aren't fighting a losing battle against an AI that has significantly better reaction time than them.

AA2 was actually sorta neat for this, the AI was quite well detailed in that game and there were some great stories coming out because of that.
Too bad Illusion went the "just make it a glorified poser" way now…

...

I wonder if that works in other games?

Still, actual fucking in h-games is just playing a stiff 2-second animation, maybe with a few gimmicks thrown in.
How are we supposed to make our fuckable and lovable robot waifus work if we can't even get it right in VR?

I wish people would stop using the term "AI" incorrectly. Actually artificial intelligence does not exist, has never existed, and likely will not exist until scientists figure out how the human brain works and can replicate it using an artificial medium. Right now, all we have are computers and bigger computers. They are tools with some level of autonomy but have ZERO intelligence. So the only way to program a computer player is to either make them hilariously incompetent or give them literal hax to overcome that incompetence.

The term doesn't mean the same under all contexts, dood.
When talking about videogames, AI is just the script that chooses the actions for any NPC and that's it.

I think predictable and exploitable AI is better for fast paced games, but you can't deny that extremely smart and realistic AI would be better for stuff like horror games.

I thought that was how all AIs worked, or at least that's what I'd do.

That's not what OP was implying. It reads like he thinks real AI is just around the corner and that it would be really cool to have them in games.

deepmind.com/blog/deepmind-and-blizzard-release-starcraft-ii-ai-research-environment/
Fortunately for you Google decided to get their DeepFreeze SC2 account.
If anything it will be interesting to see how long it will take to just use some cheese.

I don't understand what that means. What is punching in black death frame? I have also heard of some people making AI for Starcraft and having competitions with them as well. It appears the eastern Europeans are on the case here: sscaitournament.com/


They seem to have a lot of resources on it even live streaming a match between the custom AI. That was an interesting point that an user made here , that games with strong AI would only be made for spectating entertainment. Why would that be?

Because the interesting thing with a "learning AI" isn't so much the player and what he does but rather how the AI adapted to him and what's it gonna do next.

Let's say you have an horror game with a monster that evolves with time. It will constantly try to kill you but he will escape when low on health.
What the player does is interesting to some degree, but in the end you're just trying to survive, shooting at it, setting traps, gathering resources, pretty default stuff.

But if the monster adapts and changes strategy based on what you previously used, that's amazing even if done in a simple way.
Let's say you use a flamethrower, heavily scorch him several times. When he comes next holding a table or some other shield to block the flames, that's gonna impress the viewers far more than your flamethrower.


There's also the fact that, at some point the behaviour of the AI becomes so complex that it's actually interesting to observe it, maybe more than the player itself.
If you had a game like The Ship where everyone has a target they have to kill but the AI actually made alliances, pacts, trades and used diplomacy amongst itself and with the player, it would be far more entertaining to see the plans they come up with then seeing the player just killing people.


Or an even more simple example: imagine The Sims, except it doesn't suck.

Prinny is so fucking adorable.
I just wanna fiddle with his little peen and watch his flustered reactions.

that's pretty gay user

Isn't the hard part about making good AIs programming them rather than processing power?

...

Mix of both really.
As there's always a budget consideration for each frame's update time, and if you want to run at 60fps; that budget is 16.6ms per frame.

Good, complex AI is admittedly a challenge to program (though, there certainty isn't a lack of resources, it just takes time/dedication), and making a sufficiently complex AI efficiently fit within said budgeted resources in a strained update loop is doubly difficult (thus, many are cut in complexity, in favor of using fewer resources per update).
However, access to GPU async compute can be beneficial for this… that is, if one's algorithm is suited for a SIMD (single instruction (a single main function), multiple data (can utilize (also output) multiple sets of data)) algorithm structure that async compute utilizes; in addition to being able to mitigate the GPU -> CPU transferring time w/an efficient transfer scheme (akin to consideration of a per frame CPU "update loop", but a GPU->CPU update loop)… in addition to making it at a scale that makes the additional computational power of the GPU worth this overhead (akin to CPU multithreading overhead, and considerations; although, for GPU async compute we're thinking in much larger scales for threads in parallel).

So it's a nice consideration that GPU async compute could be used, but it's has a whole load of problems to solve for each particular use; in addition to not necessarily being usable for the "standard" scaled/complexity of AI approaches.

Also, OP was referencing the usage of GPU compute for neural networks (massively scaled training, which is suited quite well to SIMD algorithms, as u can easily vectorize the matrix tables generally used in neural networks weights, plus propigating changes w/gpu scatter/etc), which are generally used in an application that doesn't require real-time updates; nor does it have to sync with an entire game world's CPU side update loop.
So… while the use of GPU async for AI isn't "new", it certainly hasn't been utilized enough in a game setting to be considered viable from a consensus point of view.

This is why I hate armchair autists.

In fairness, that AI doesn't look like it's playing against actual people.

...

That's a specific network built for that level and that level alone. Stick it in another level and it'll be useless. I'm not saying we won't get to the point where a network will be able to learn as it goes or some shit, but we aren't there yet.

Fear's AI is a perfect example of emergence. the AI of individual soldiers in Fear is actually very simple, but in-game they all interact with each other in such a way that they create the illusion of being smarter and more complex than they actually are.

Except you're wrong. Sethbling gave the AI no help in learning how to play the level and at no point does he say anything that suggests it wouldn't work on another level.

Try reading what I wrote. THAT network will not work for other levels, because it is trained for that level alone. If you train it for a different level, it will no longer be able to complete the first. Perhaps if you ran each levels in a row, which may as well be one giant level.

It's not a generalized AI for completing levels. It's a network for completing that specific level with that specific set of obstacles/

then keep the fittest species for each level

...

This is the keywording here.

That neural network did utilize training, but, "unsupervised training" in AI lingo (as compared to supervised training); in essence, all neural networks require some kind of training.

Now… I want to get the point across on why it will not perform the same on level 2, as it did on level 1 (i.e. learning from lvl 1, and beating lvl 2 after a few tries); which is due to the approach used for training this AI.
So… this person trained the AI via an evolution algorithm, and required many… many generations (basically, iterations of groups, that have the highest "fitness" level, in this case, I think the metric was how far they got in the level) to reach a sufficient level of competence for a singular level.
Now, it's trained for that specific level, and didn't really "know" what to do situations; it's an approach that makes this an extremely dumb AI that has no "real" spatial awareness, and no "real knowledge" of enemies or their locations, and only learned through gaining a "higher fitness" level, and thus eventually "evolving" through generations to beat the level.
Now, the reason that this is important, is because when you put it into another level; it will be dumbfounded, and generation after generation will make the same exact mistakes they made on the first level it was trained on.
Now, you would think it would be able to beat it after a few tries; due to beating level one, but, the way it learned wasn't through absorbing it's mistakes via a reward metric (this enemy, represented by a moving point, is bad, avoid it to improve these "reward" metrics); it learned via trial and error (jump at this point (which avoids the enemy, the AI does NOT know the reason, only to jump), due to that being a factor in it having a fitness level high enough to proceed to the next generation).

A way better approach is the vid linked in this user's post:

Which can pick up a different game, and figure out what to do based on what appears to be a method utilizing a reward metric (or called, reward learning, or reinforcement learning; like how our brains generally learn due to "reward networks" in our brain, f.e. fapping is good, spread your genetic line).

A very early, and currently not very computationally efficient approach to AI (compared to specialized approaches, at least in the game dev scene… also the inefficiently is mostly due to recursion… which could be mitigated with a more "recursive specialized" language); although, it's a "jack of all trades" which is extremely useful.
However, being a "jack of all trades" does have its down sides… as specialized algorithms for particular tasks can be way more efficiently accomplished by dedicated AI to this task.
Although, there is certain things that are way easier to do with neural networks compared to specialized AI.
So it's really a tossup, of whether or not one has the ms budget (games are… a touchy subject here, as mentioned in my previous post), and if one can train the neural network sufficiently; to perform at the specified level of competence.

AI in the sense of modern pattern recognition software is useless for vidya.
a well written finite state machine gives the results you want, which is the appearance of intelligent play. the problem is writing good criteria and choices for the fsm to make.

that's a lot of words to say "you could spend a lot of time and money to write a complex AI to play Quake that would get stomped by a Quake bot written decades ago", but basically true.

(checked)
Well, honestly it hasn't been done really well yet; from a consensus point of view you're 100% right though, as anyone purporting otherwise has almost no evidence saying otherwise.
Although, neural networks are a "jack of all trades", and could definitely be shaped into fulfilling a role that's it's well suited for; although as mentioned in my first post there's a lot of potential issues here, and the undeniable point that there's more suited -specialized- approaches for the requirement of games.

agreed, games definitely benefit from more specialized approaches; due to the plethora of issues that arise such as budget, scale, and tying anything more complex into the entirety of the game.
Honestly, halo:CE's behavior trees we're goddamn amazing, and still blow me away to this day at how well they crafted those behaviors way back in 01'.

(checked)
hah, yeah I try to explain all the details so there's no misinterpretation of the points I was making; as that seems to happened a lot on the internet in general.
I'm generally a "long winded" person so this is the usual format of my posts heh.

I'm not familiar with quake bots, if they cheated or not, or the approach they used (I prefer unreal tourney myself); although, a sufficiently trained neural network (with the right approach) could, at times, convincingly "mimic" the action of a human playing quake (as the unreal neural network bot achieved, "passing" unreal's version of the "turing test" for fooling people into thinking the bot was a human for ~50% of the time).

more complex AI just means that more complex actions can be executed. Which means more ways that they can be abused. So I'm confident I'll be able to continue my cheesey ways far into the future.

I dunno about that. Something patent in [Bethesda Game] is that a very effective strategy against any non-ranged enemy is to find a small wall that you can jump over and keep switching sides, prompting the melee enemy to run at you when you're on their side and running away when you're on the other side.
So Holla Forums's practically Bethesda AI-tier, lol
If at any point the AI was capable of understanding what jumps actually are, or at least to set up ambushes or properly flank, the game would certainly be much harder since the AI would cheese you instead.

I remenber killing a Deathclaw in FO4 very easily with this trick, which made me glad I pirated the game instead, and then there was those tricks about staying inside small cafes where a Deathclaw can't go inside but you can shoot from the window, and the dumb thing never things about moving to cover or wait for you outside instead or just give up and go away.

The problem with AI in video games is the number of people who legitimately want to play against opponents that learn and adapt but are too autistic to just go online and play humans is rather small. What most people want is AI that makes them, the player, feel smart. Which means mostly being dumb and predictable without looking dumb and predictable. So it ends up being a design problem rather than an AI problem.

They won't, as a general rule, outside of niche stuff like the Creatures games.
The thing about convolutional neural nets, genetic algorithms, and the like, is that they are best suited to situations in which you have a desired outcome that is relatively easy to measure, but you aren't certain what the most effective behavior is to achieve that outcome, or alternately, how to program it. Hence, things like winning at Chess or Super Mario, which have extremely simple end goals (checkmate the opponent, move right), and not terribly complicated "mid" goals (trade a weak piece for a strong piece, jump on a platform that isn't to the right but takes you to a pathway that is). These methods are designed to solve a problem, with an extremely large amount of flexibility in approach, relative to other AI methods. The tradeoff being that you largely, intentionally, give up control of that behavior. NNs and GAs often generate solutions that are difficult for the people who created them to quickly understand and therefore alter. Sure, you can rerun the network with different training parameters to tweak your solutions that way, but you don't know what your actual results will be or how many times you'll have to change things to get the behavior you want. After all, if you did, you would just be programming that in the first place.

What most people who post AI threads ultimately describe wanting is not AI that is "better" in terms of simply winning, but in terms of simulating specific behavior. In particular, AI dreamers usually describe more complex interactions between AI characters themselves, and the simulated world they live in, like advanced diplomacy and believable dialogue. And the thing is, DNNs are absolutely the WORST options for attempting to simulate a specific behavior. The whole point of using them is that they will frequently do something you completely don't expect, and that's untenable from a game design and maintenance standpoint.

What you should actually be hoping for is progress in terms of a mathematical understanding of complex interactions in sociology and zoology, because just as it would be basically impossible to program an accurate physics simulation without some kind of basic mathematical and logical understanding of physics, it's going to remain impossible to program complex behavioral simulations as long as they are largely dominated by evolutionary psychology and the like. Progress is being made, mind you, just not very quickly.

on that topic im pretty impressed with the minigames with the bandit in yoshi's island on the SNES. the seed spit/coin collect etc. games

But what if the goal is made more complex, though? What if it's calculated with a different formula that takes different parameters and the best result is attainable only by learning and using multiple things?

The Mario example is indeed very simple since the AI accrues points simply by how far it made it to the right.
But what if there was another parameter that contributed to fitness, like the points it made or how many coins it collected?
Simply reaching the end would not be the best possible scenario and AIs that also managed to score points or collect coins on the way there would actually have higher fitness, but if they didn't made it to the end or at least far enough, it wouldn't really count either.

For instance a simple equation of X = (A * B)/(A + B) would make it so both parameter A and B are equally important and focusing too much on either gives you sub-optimal resources.
I'm sure this would imply several more generations for the simulation to run so it can also take that parameter, and a more advanced equation would be needed if there were more than 2 parameters…
But wouldn't this solve this problem?


Specifically for the idea of "advanced diplomacy and believable dialogue", this is no different, the AI just needs a different parameter for Fitness depending on it's personality. If you have an actor that's greedy, anything that rewards him with money should increase his fitness, if it's a religious actor, everything related to spreading and strengthening his religion should increase it's fitness.

And even there, you can have 2 different fitness algorithms running, one that tries to conserve the actor alive for as long as possible and spread his genes as much as it cans, while another acts on it's free time to advance it's specific behavior and train him in how to do that even better.

This way you could have a Paladin that travels the land to smite evil. His first concern is to stay alive, eat and rest, eventually finding a wife to sire his children, the longer he stays alive and the more children he breeds the better "survival fitness" he accrues when he finally expires.
But whenever his needs are met and it's not possible to breed more, he'd try to find evil-doers and kill them, gaining "paladin fitness" for every one it manages to kill.

You could then re-use his training for other people, with the "survival fitness" being suited for anyone that just wants to live with the "paladin fitness" being applicable as a side-routine to anyone that wants to opose evil.

Having the technology doesn't mean that developers will utilize it, the limits of current hardware is very rarely pushed.

Also functional AI that beats humans at every task will be the death of human spirit and will cause a downward spiral of degeneracy and gluttony

What on earth makes you think AI is going to get really good soon?

The trash tier shit on the shelves at the moment betrays the reality of it, I'm sad to say you're in for a sorry disappointment m8, at best you'll get stuff thats massively scripted to *not* look scripted or else garbage tier kid-friendly bullet magnets for the foreseeable future.

The industry needs to be crashed into a wall before anything serious will be made.